Mindspeed Mm Env Setup

MindSpeed-MM Ascend NPU Base Environment Setup

This skill guides users through setting up the base environment for MindSpeed-MM multimodal training on Huawei Ascend NPU.

Important: This guide only covers the base environment. Different multimodal models (qwen3vl, wan2.2, hunyuanvideo, etc.) have vastly different dependency version requirements that may conflict with each other. Model-specific dependencies must be installed on top of the base environment. After completing this guide, refer to the corresponding model's SKILL for additional configuration.

Component Relationship

Megatron-LM (NVIDIA)     <- Distributed training core (TP/PP), uses core_v0.12.1 branch
    |
MindSpeed (Huawei)       <- Ascend adaptation layer, monkey-patches Megatron kernels
    |
MindSpeed-MM (Huawei)    <- Multimodal application layer: VLM/generation/omni-modal training

MindSpeed-MM shares the underlying dependency stack (CANN, torch_npu, MindSpeed, Megatron-LM) with MindSpeed-LLM, but targets multimodal scenarios at the application level (vision-language models, video generation, speech synthesis, etc.).

Quick Start -- 6 Steps to Complete the Base Environment

# 1. Activate CANN environment (CANN 8.5.0+ path; for older versions use /usr/local/Ascend/ascend-toolkit/set_env.sh)
source /usr/local/Ascend/cann/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh

# 2. Install PyTorch + torch_npu — USE PRE-BUILT WHEELS from https://gitcode.com/Ascend/pytorch/releases
# Do NOT use `pip install torch==2.7.1` — aarch64 wheels on PyPI are unreliable
# For Python 3.10 + aarch64:
pip install torch-2.7.1-cp310-cp310-manylinux_2_28_aarch64.whl
pip install torch_npu-2.7.1rc1-cp310-cp310-manylinux_2_28_aarch64.whl
pip install numpy pyyaml scipy attrs decorator psutil

# 3. Clone and install MindSpeed (pinned to a specific commit)
git clone https://gitcode.com/ascend/MindSpeed.git
cd MindSpeed
git checkout 93c45456c7044bacddebc5072316c01006c938f9
pip install -r requirements.txt && pip install -e .
cd ..

# 4. Clone Megatron-LM and copy the core module into MindSpeed-MM
git clone https://github.com/NVIDIA/Megatron-LM.git
cd Megatron-LM && git checkout core_v0.12.1 && cd ..
git clone https://gitcode.com/Ascend/MindSpeed-MM.git
cp -r Megatron-LM/megatron MindSpeed-MM/

# 5. Install MindSpeed-MM
cd MindSpeed-MM
pip install -e .

# 6. Verify the environment
python -c "
import torch
import torch_npu
print(f'NPUs available: {torch_npu.npu.device_count()}')
print(f'NPU ready: {torch.npu.is_available()}')
"

One-Click Install Script (Qwen3/Qwen3.5 ONLY)

MindSpeed-MM provides scripts/install.sh, but official docs state it only fully supports Qwen3 and Qwen3.5 models. For all other models (Qwen2.5VL, Wan2.1/2.2, InternVL, HunyuanVideo, etc.), use the manual install above.

# ONLY for Qwen3 / Qwen3.5:
git clone https://gitcode.com/Ascend/MindSpeed-MM.git
cd MindSpeed-MM
bash scripts/install.sh --msid eb10b92 && bash examples/qwen3_5/install_extensions.sh

Warning: On ARM (aarch64), install.sh uses pip install torch torchvision torchaudio from PyPI, which frequently fails because aarch64 wheels are not consistently available. For ARM, prefer the manual install method with pre-built wheels from Ascend PyTorch Releases.

install.sh Parameter Reference

Example: Specify PyTorch 2.6.0 and auto-confirm

bash scripts/install.sh --msid eb10b92 -t 2.6.0 -y

Version Compatibility Matrix

The above reflects the current master branch. Both MindSpeed-MM and MindSpeed are under rapid iteration; it is recommended to use a pinned commit rather than master HEAD. Check CANN version: cat /usr/local/Ascend/cann/latest/aarch64-linux/ascend_toolkit_install.info (CANN 8.5.0+) or cat /usr/local/Ascend/ascend-toolkit/latest/aarch64-linux/ascend_toolkit_install.info (older versions).

Docker Container Creation

docker run -it --name mindspeed-mm \
    --privileged \
    --network host \
    --ipc=host \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64:/usr/local/Ascend/driver/lib64 \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /home/workspace:/home/workspace \
    <cann-image> bash

Warning: If --ipc=host is not available, set --shm-size=16g. Without proper shared memory configuration, DataLoader workers will crash with Bus error. As a workaround, set --num-workers 0 in all training and feature extraction scripts.

Note: The CANN container image does not include ML dependencies; after entering the container you must manually install PyTorch, torch_npu, etc. It is strongly recommended to create a separate container for each multimodal model to avoid dependency version conflicts.

Directory Structure

After installation, the workspace structure should look like this:

workspace/
├── MindSpeed/                    # Acceleration library
├── MindSpeed-MM/                 # Main project
│   ├── megatron/                 # Core module copied from Megatron-LM
│   ├── mindspeed_mm/             # Multimodal core code
│   ├── examples/                 # Example scripts and configs for each model
│   ├── scripts/
│   │   └── install.sh            # One-click install script
│   └── pyproject.toml            # Base dependency definitions
├── Megatron-LM/                  # NVIDIA original repo (used only to copy megatron/)
├── model_from_hf/                # Model weights
└── dataset/                      # Training datasets

Base Dependency Versions (pyproject.toml)

The base installation (pip install -e .) installs the following core dependencies:

Model-Specific Dependency Conflict Warning

Critical Warning: The multimodal models supported by MindSpeed-MM have vastly different dependency version requirements that may conflict with the base environment.

Known conflict examples:

• qwen3vl: Requires installing transformers from git source (specific commit c0dbe09), incompatible with base version 4.57.0
• wan2.2: Requires diffusers==0.35.1, peft==0.17.1, overriding base versions
• hunyuanvideo_1.5: Requires transformers==4.57.1, diffusers==0.35.0, also incompatible with wan2.2
• deepseekocr: Requires transformers==4.46.3 (downgrade!), conflicts with all other models

Docker shared memory: When running in Docker containers, ensure adequate shared memory is configured (--ipc=host or --shm-size=16g). Default Docker shm (64MB) will cause Bus error in DataLoader workers. If you cannot change the container settings, set --num-workers 0 in all training scripts.

Isolation recommendations:

• Create a separate Docker container or conda/venv environment for each model
• Complete the base environment setup first, then install model-specific dependencies in an isolated environment
• See references/version-matrix.md for the full conflict matrix

Environment Verification Checklist

Smoke Test

Run this quick preflight check before starting any training job:

python3 -c "
import torch, torch_npu
assert torch.npu.is_available(), 'NPU not available'
print(f'NPU count: {torch_npu.npu.device_count()}')
import transformers, diffusers, mindspeed
print(f'transformers={transformers.__version__}, diffusers={diffusers.__version__}')
print('Smoke test passed')
"

If any import fails or the assertion triggers, revisit the corresponding installation step above.

FAQ

Q: ModuleNotFoundError: No module named 'yaml' when importing torch_npu

The CANN Docker image is missing basic Python packages:

pip install numpy pyyaml scipy attrs decorator psutil

Q: pip install or git clone fails due to network issues

Use a proxy:

export http_proxy=http://127.0.0.1:7890
export https_proxy=http://127.0.0.1:7890
pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cpu

Q: Cannot find the Megatron-LM checkout branch

Make sure you are using the correct branch name (note the core_v prefix):

cd Megatron-LM
git branch -a | grep core
git checkout core_v0.12.1

Q: MindSpeed commit does not exist

Make sure you are using the full commit hash:

cd MindSpeed
git log --oneline | head -20
git checkout 93c45456c7044bacddebc5072316c01006c938f9

Q: pip install -e . fails when installing MindSpeed-MM

Check that the megatron module has been copied into the MindSpeed-MM directory and that pyproject.toml exists:

ls MindSpeed-MM/megatron/
ls MindSpeed-MM/pyproject.toml

Q: Training fails with Communication_Error_Bind_IP_Port

A stale process from a previous run is holding the port. Kill it and use a different port:

# Find and kill stale processes
ps aux | grep torchrun | grep -v grep | awk '{print $2}' | xargs kill -9
# Or use a different port in the training script
export MASTER_PORT=6100

Q: Base environment is broken after installing model-specific dependencies

This is expected behavior. Different models require conflicting versions of transformers/diffusers/peft. Solutions:

# Option 1: Separate Docker containers (recommended)
docker run -it --name mindspeed-mm-qwen3vl ...

# Option 2: conda virtual environments
conda create -n mm-qwen3vl python=3.10
conda activate mm-qwen3vl
# Re-run base installation + model-specific dependencies

Next Steps

The base environment setup is complete. Next, install additional dependencies for the specific model you need. Refer to the corresponding model's SKILL.

References

• Version Conflict Matrix - Dependency version conflict details for each model
• MindSpeed-MM Repository
• MindSpeed Repository
• Megatron-LM
• Ascend Documentation